Banking on Trust: Debit Cards, Cash Transfers, and Savings in Mexico

Transcription

1 Banking on Trust: Debit Cards, Cash Transfers, and Savings in Mexico Pierre Bachas UC Berkeley Paul Gertler UC Berkeley Enrique Seira ITAM December 21, 2015 Sean Higgins Tulane University Abstract Trust is an essential element of economic transactions. Trust in nancial institutions is especially low among the poor, and may explain in part why the poor do not save formally. Debit cards provide not only easier access to savings (at any bank's ATM as opposed to the nearest bank branch), but also a mechanism to monitor bank account balances and thereby build trust in nancial institutions. We study a natural experiment in which debit cards were rolled out to beneciaries of a Mexican conditional cash transfer program, who were already receiving their transfers in savings accounts through a government bank, but were choosing not to use their accounts to save. Using 3 years of administrative data on transactions and balances in over 300,000 bank accounts, we nd that after receiving a debit card, the transfer recipients do not increase their savings for the rst 6 months, but after this initial period, they begin saving and their marginal propensity to save increases over time. During this initial period, however, they use the card to check their balances frequently; the number of times they check their balances decreases over time as their reported trust in the bank increases. Using household survey panel data, we nd an increase in overall savings, rather than substitution from other forms of saving; we also nd that consumption of temptation goods (alcohol, tobacco, and sugar) falls, providing evidence that saving informally is dicult and the use of nancial institutions to save helps solve self-control problems.

2 Virtually every commercial transaction has within itself an element of trust, certainly any transaction conducted over a period of time. It can be plausibly argued that much of the economic backwardness in the world can be explained by the lack of mutual condence. Kenneth Arrow (1972) 1 Introduction Trust is an essential element of economic transactions. This is especially true for savings, where the transaction takes the form of a promise of future returns. Trust in nancial institutions, however, is low worldwide: in 40 percent of countries included in the World Values Survey, the majority of people report low trust in banks (Figure 1). In Mexico, where 54 percent of people report low trust in banks, trust is especially low among the poor: 71 percent of those with no formal education report low trust in banks, compared to 55 percent of those who completed primary school and 46 percent of those who completed university (Figure 2). 1 The lack of trust in nancial institutions in Mexico is not unfounded: over the last 15 years, there have been numerous highly publicized frauds in Mexico where poor savers have lost the money they deposited in nancial institutions. 2 Furthermore, bankers in Mexico loot money by directing a large portion of bank lending to related parties, i.e. shareholders of the bank and their rms (La Porta et al., 2003). When contract enforcement is low (as it was shown to be after some of these frauds), trust and trust building become more important (Karlan et al., 2009), and people are understandably reluctant to leave their money with nancial institutions when trust is low: Guiso et al. (2004) nd that lower trust is associated with holding more cash on hand. Along with transaction costs, low trust is frequently listed by the poor as a primary reason for not saving in formal savings accounts when these are explicitly oered to them. For example, in a follow-up survey to a randomized experiment, Dupas et al. (forthcoming) nd that families cite unreliability and risk of embezzlement as two of the primary reasons for not opening a savings account after the opening fee is waived. 1 Low trust in banks is dened as not very much condence or none at all for the item banks in response to the following question from Wave 6 of the World Values Survey: I am going to name a number of organizations. For each one, could you tell me how much condence you have in them: is it a great deal of condence, quite a lot of condence, not very much condence or none at all? In the Spanish version of the survey, the word conanza is used, which directly translates to both condence and trust. 2 We scraped the web for Mexican newspaper articles describing such frauds and found hundreds of articles each year. For example, one article is titled Senator estimates frauds against savers exceed 25 billion pesos. It is also telling that articles with nancial advice in Mexico have titles like How to Save for Your Graduation and Avoid Frauds and Retirement Savings Accounts, with Minimal Risk of Fraud. 1

3 We study a natural experiment in which debit cards were rolled out to beneciaries of the Mexican conditional cash transfer program Oportunidades, who were already receiving their transfers in savings accounts through the government bank Banse, but were choosing not to use their accounts to save. 3 Debit cardsand mobile banking more broadlycan alleviate the low-trust constraint by providing not only easier access to savings (at any bank's ATM as opposed to the nearest bank branch of a particular bank), but also a mechanism to monitor bank account balances and thereby build trust in nancial institutions. The picture that emerges from our ndings is that the debit card decreases the cost of monitoring and using the account, which leads to monitoring the bank through balance checks, increased trust in the bank, and in turn, a gradual increase in overall savings. Saving in the account appears more eective than attempting to save with the cash in hand or under the mattress, in which case the money is hot (Ashraf, 2009, p. 1267): indeed, beneciaries with debit cards decrease their spending on temptation goods, suggesting that self-control problems prevented them from saving as much as they wanted informally, and that saving formally in trusted bank accounts can help solve these self-control issues. This is the rst paper we know of that shows how the convenience and monitoring capacities aorded by debit cards aect trust, the use of formal savings accounts, and overall saving. Previous papers have focused on small-scale interventions looking mostly at the eect of changing fees/prices or oering commitment devices on account opening and use; 4 indeed, Banerjee (2013, p. 488) notes that even though both microsavings and micro-insurance are potentially important parts of the overall agenda, the research on these services is still in its infancy. Furthermore, the policy experiment we study is particularly interesting for several reasons. First, while most papers on the causal eects of various savings interventions rely on small-scale experiments, we study a nationwide (and, we argue, exogenous to saving) expansion of debit cards aecting hundreds of thousands of households living in 275 dierent urban localities. Second, it uses a standard technologythe debit cardand therefore should be scalable to millions of government cash transfer recipients worldwide. 5 3 The program began as Progresa in rural areas and was later renamed Oportunidades and expanded to urban areas. It has since been renamed Prospera, but we use the name that was in place during the time of our study, Oportunidades. 4 Schaner (2015a) analyzes the eect of ATM cards, but in her study the ATM cards also reduce withdrawal fees which are charged not only by ATMs but also by bank tellers in her studyby about 50 percent, which prevents her from disentangling the eects of the reduced direct transaction costs (lower withdrawal fees) from that of indirect transaction costs and trust building. 5 Many countries, such as Brazil, Colombia, Guatemala, and Indonesia, are already digitizing their government to person payments, with little knowledge of what binding constraints prevent the poor from saving, and which policy 2

4 Third, we use a rich combination of administrative data on account balances and transactions, household survey panel data, and cross-sectional data to study formal and overall savings, account transactions such as withdrawals and balance checks, trust, learning the technology, and changes in consumption. Unsurprisingly, trust of the bank among these cash transfer recipients was initially low. Only 16% of beneciaries with bank accounts reported they were safe or trustworthy. When asked why she did not save in her account shortly after receiving her card, one beneciary responded that it is wellknown that if I leave part of the money on the card, they take it from me or it no longer appears in the account. 6 Another beneciary described the bank accounts as a technology imposed on us that scares me, and makes me feel frustrated and anxious (Inter-American Development Bank, 2010). Government ocials administering the Oportunidades program recount anecdotes of beneciaries depleting the money left in their account by repeatedly checking their balances (balance checks are charged a fee by ATMs in Mexico), and we indeed see in the administrative account transaction data from Banse that some Oportunidades beneciaries check their balances up to 40 times per two-month payment period, paying a total cost of over 10 percent of average benets received from the program. During the two years after receiving the debit card, savings in the account increases substantially. The increase is gradual, with no eect for the rst 6-8 months in the rst wave of the roll-out and an increasing propensity to save over time in both waves. Combining our results from the accounts with household survey data on income, we estimate that after one year with the card, the percent of total income saved increases by 5 percentage points relative to the control group; after two years, those with cards save 10 percentage points more each payment period. This is in itself an important nding, as it shows that there was unmet demand for formal savings among these families, but that some form of constraint was preventing them from using their Banse accounts to save. The delayed eect and temporal gradation of the increase in the propensity to save suggests some kind of learning. We explore two kinds of learning that may be occurring: the rst, which we call learning to use or operational learning, involves getting more acquainted with how the account, debit card, and ATMs work; the second, which we call learning to trust or building trust, interventions can relax these constraints. 6 Response from the Encuesta de Características Sociodemográcas de los Hogares Urbanso (ENCASDU) 2010, described in more detail in Section 3. 3

5 involves updating beliefs about the trustworthiness of the bank and the likelihood of losing the money in the account. Using a survey conducted by Oportunidades and exploiting the variation in the amount of time with a debit card (exogenously determined by the locality-level roll-out), we nd that very few beneciaries (about 1 percent) report not saving in the account because they do not know how to use the technology, and that this proportion is constant over time with the debit card. On the other hand, 27 percent of beneciaries that have had the card for 6 months or less report not saving because they do not trust the bank, compared to 17 percent of beneciaries who have had the card for more than 6 months. We use a second survey conducted by Oportunidades to investigate mechanisms behind these two potential forms of learning, and nd that those who have had the card for at least 6 months report checking their balances signicantly less frequently and making fewer trips to the ATM solely to check their balances (without withdrawing money), the mechanism they use to monitor the bank and build trust once they have the card. Beneciaries with more than 6 months with the card check their balances without withdrawing money 36 percent less frequently than those with less than 6 months of experience with the card. Using administrative data on balance checking, we nd the same trend as in the self-reported survey data: when beneciaries get the debit card, they check their balances often, but graduallyas trust is builtthey check less. Meanwhile, selfreported knowledge of how to use the technology remains fairly constant over time, with statistically insignicant dierences in the proportion reporting that it is dicult to use the ATM, that they get help using the ATM, or that they know they can save in the account. Given the increase in savings in the Banse account, we then test whether this increase is an increase in total savings or a substitution from other forms of saving (e.g., from under the mattress, rotating savings and credit associations [ROSCAs], or other bank accounts). Because the debit cards make saving in the Banse account more attractive, beneciaries could merely be substituting Banse savings from other forms of saving. Using survey data on consumption, income, and assets, we nd that the treatment group increases savings by about 5 percent of income relative to the control group after having the card for nearly one year, which is similar in magnitude to the eect we see in the administrative bank account data. We nd no dierential change in the income, purchase of durables, or stock of assets in the treatment group compared to the control, but a decrease in consumption, suggesting that the increase in bank savings represents an increase in overall savings 4

6 rather than savings shifting, and comes from a voluntary decrease in current consumption. Moreover, we nd that the categories of consumption with a statistically signicant decrease are alcohol, tobacco, and sugarthe most frequently mentioned temptation goods in Banerjee and Mullainathan (2010)and transportation. This provides evidence that informal savings is dicult and the use of nancial institutions to save solves self-control problems. Given our results, government cash transfer programs could be a promising channel to increase nancial inclusion and enable the poor to save, not only because of the sheer number of the poor that are served by cash transfers, but also because many governments are already embarking on digitizing their cash transfer payments through banks and mobile money, and because the technology of debit cards and ATMs is simple and prevalent. 2 Context of the intervention Half of the world's adults do not use formal nancial services such as savings accounts. The poor disproportionately lack access, with nearly 90 percent of the unbanked living in developing countries (Chaia et al., 2009). Although individuals may optimally choose not to save in formal accounts even if provided with aordable and accessible savings products, several recent studies have shown that there is unmet demand for formal savings products among the poor, and that providing access to such products can have a robust positive eect on welfare proxies. 7 Although the poor do save via cash at home, this cash may be susceptible to theft (Wright et al., 2014), temptation spending (Ashraf, 2009), or pilfering by friends and relatives who request money (Baland et al., 2011; Dupas and Robinson, 2013b). The poor can also save by investing in assets such as livestock, but saving in these assets is illiquid and requires lumpy investments (Fafchamps et al., 1998). If saving informally is dicult, there may be unmet demand for formal savings products among the poor, andunder the right conditionsadding formal accounts as another savings product in their portfolio could lead to increases in total savings and welfare. But the solution is not just to open bank accounts for the poor. Governments have recently given savings accounts to millions of poor beneciaries by depositing cash transfers directly in bank 7 Being oered a savings account increased business investment by 46% and daily private expenditures by 38% for micro-entrepreneurs in Kenya (Dupas and Robinson, 2013a), increased agricultural output by 15% and household spending by 11% for tobacco farmers in Malawi (Brune et al., forthcoming), and increased total assets by 12% and expenditures in education by 20% in Nepal (Prina, 2015). 5

7 accounts, only to nd that most of these accounts are inactive, other than one withdrawal per pay period used to withdraw the entire transfer. Academics have also reached this conclusion: when they have oered savings accounts to poor individuals and subsidized the minimum balance requirement, opening fee, or withdrawal fees, they nd that most accounts that were opened have less than two transactions or deposits after the rst six months or one year (e.g., Ashraf et al., 2006; Dupas and Robinson, 2013a; Dupas et al., forthcoming; Karlan and Zinman, 2014; Schaner, 2015b). 8 We examine the case of the highly studied conditional cash transfer program Oportunidades and, similarly, nd that savings accounts were barely used prior to the introduction of debit cards. At baseline, the average number of deposits per bimester was 1.05 (including the deposit from Oportunidades), the average number of withdrawals was 1.02usually cashing the entire transferand the average bimonthly account balance was 618 pesos, which arises almost entirely from the mechanical eect on average daily balance over the bimester of not withdrawing their transfer immediately. 9 Indeed, prior to receiving cards, 98.9 percent of the transfer was withdrawn on average during the rst withdrawal following payment. We exploit the gradual roll-out of debit cards to urban beneciaries of Oportunidades, a conditional cash transfer program that began in rural Mexico in 1997 under the name Progresa, and expanded to urban areas beginning in Oportunidades is possibly the most well-known conditional cash transfer program worldwide, with a history of rigorous impact evaluation (e.g., Gertler, 2004; Parker and Teruel, 2005). The program provides bimonthly cash transfers to poor families in Mexico, seeking to alleviate poverty in the short term and break the intergenerational poverty cycle in the long term by requiring families to send their children to school and have health checkups. The program has had a number of benecial eects, mostly estimated experimentally using the randomized phase-in of the program in rural areas. These include increased birthweight (Barber and Gertler, 2008), reduced child behavioral problems (Fernald et al., 2009), increased time spent on homework (Behrman et al., 2012), increased years of schooling in both the short term (Schultz, 2004) and longer term (Behrman et al., 2011), increased productive investments (Gertler et al., 2012), decreased depression among recipient mothers (Ozer et al., 2011), and reduced elderly 8 One exception is a study in Nepal by Prina (2015), where the account had no fees and households who were oered accounts lived in close proximity of the bank; in this case, 80 percent of individuals oered an account made at least two deposits in the rst year. 9 The program is paid in two-month intervals, which we refer to throughout the paper as bimesters. (The Spanish word bimestre is more common than its English cognate, and is used by Banse and Oportunidades.) 6

8 mortality (Barham and Rowberry, 2013). Today, nearly one-fourth of Mexican households receive benets from Oportunidades (Levy and Schady, 2013). After expanding to urban areas, Oportunidades paid benets to some urban recipients through bank accounts. The original motives for paying through bank accounts were to decrease corruption (automatic payments through banks would lower both the ability of corrupt local ocials to skim o benets, and of local politicians to associate themselves with the program through face-to-face contact with recipients when they received their transfers), decrease long wait times for recipients (who previously had to show up to a payment table on a particular day to receive their bene- ts), decrease assaults on program ocers and recipients transporting cash on known days, and increase the nancial inclusion of poor households. By the end of 2004, over one million families received their benets in savings accounts through two banks partnered with the program: Banse, a government bank created to increase savings and nancial inclusion, and Bancomer, a commercial bank (Figure 3). After diculties working with a commercial bank, Oportunidades phased out the accounts through Bancomer and changed them to Banse accounts by mid Initially, this account had limited attractiveness as a savings product: it paid between 0.09 and 0.16 percent nominal interest per year (ination in Mexico had been close to 5 percent per year during our sample period) and had no debit card that could be used at ATMs or to pay electronically at stores, which meant that transactions had to be done at the Banse branches or, in some cases, temporary points of payment set up by Oportunidades and Banse. Because there are about 500 branches across the country, many beneciaries live far from their closest Banse branch. On the plus side, the account did not require a minimum balance. In 2009, the Mexican government announced that all recipients would receive their payments through a plastic card by In urban localities that were included in Oportunidades and in which beneciaries already received their benets in Banse savings accounts, 10 Visa debit cards would be awarded. These cards would be tied to the Banse bank accounts through which recipients already received their transfers and would enable holders to withdraw money at any bank ATM and pay electronically at any store that accepts Visa cards. In addition, the cards would include two free ATM withdrawals every bimester at any bank's ATM; after that, the fees for each ATM withdrawal vary by bank, averaging 13 pesos (about $1 using 2009 exchange rates). Of Mexico's 10 Urban localities are dened as localities with more than 15,000 inhabitants 7

9 550 urban localities, 275 were already included in the Oportunidades program and paid beneciaries in savings accounts. Of these, 143 were selected to change in 2009; we refer to this group as wave 1 of the roll-out. In this wave, approximately 100,000 Oportunidades recipients' existing savings accounts were tied to debit cards. Another wave of urban localities changed in late 2010, resulting in another 75,000 recipients receiving debit cards tied to their bank accounts; we refer to this group as wave 2. The remaining urban localities received debit cards at the end of 2011 or during 2012; we refer to this group of approximately 170,000 accounts as the control because our data period ends before they receive debit cards. Which localities switched rst was determined as a function of the proportion of households in the locality that were eligible for the program but were not yet receiving benets, since the introduction of debit cards to existing recipients was coupled with an eort to include more people as beneciaries. Table 1 compares the means of locality-level variables and account-level variables from the control, wave 1, and wave 2 localities using the population census from 2005 (the most recent year prior to the beginning of the debit card roll-out), poverty estimates from Oportunidades from 2005, Banse branch locations from 2008, and the administrative account data on average balances and transactions from Banse at baseline (the portion of 2009 before any accounts in our sample received cards). Column 6 shows the p-value of an F-test of equality of means. Because the roll-out was not random, it is not surprising that there are some dierences across treatment and control localities: treatment localities are slightly larger and beneciaries in these localities receive higher transfer amounts. The percent of the transfer withdrawn also diers (it is lower in wave 1 than the control and insignicantly dierent but with a higher point estimate in wave 2), but is high in all cases (ranging from 97.5 percent to 99.6 percent of the transfer), indicating very low savings in the account prior to receiving the card. In Section 4 we will test and show that trends of saving, income and consumption were parallel across waves. The map in Figure 4 shows that the treatment and control waves had substantial geographical breadth and that treatment and control localities were physically close. 3 Data To causally estimate the impact of debit cards on savings, we use a rich combination of data sources. First, we have administrative data from Banse at the account level. This includes information 8

10 about bimonthly average savings balances 11 for 342,709 accounts at 308 Banse branches for the period January 2009 to October Figure 3a shows the administrative Banse data relative to the roll-out of debit cards. These data also include the transfer amounts each bimester, the timing and amount of transactions made in the account, the date the savings account was opened, and the month the card was awarded to the account holder; the average account had been opened 5.3 years before getting the card. We supplement the administrative account data from Banse with household survey data merged with additional administrative data from Oportunidades. First, the Encuesta de las Características de los Hogares Urbanos (Survey of Urban Household Characteristics; ENCELURB) is a panel survey with three pre-treatment waves in 2002, 2003, and 2004, and one post-treatment wave conducted from late 2009 to early This survey has comprehensive modules on consumption, income, and assets. 12 The pre-treatment waves include a larger number of localities and households (more than 17,000 households), but due to budget constraints the wave includes 6272 households in urban and semi-urban areas (with population of 2500 inhabitants or more). We merge these data with administrative data from Oportunidades on the transfer histories for this samplewhich we use to add transfer income into total income and to identify which households are Oportunidades recipients, given the common misreporting of transfer receipt in surveys (Meyer et al., 2015)and on the dates that debit cards were distributed in each locality. Of the 6272 households in the post-treatment wave of ENCELURB, 2951 live in urban areas and are Oportunidades beneciaries when interviewed in the post-treatment wave; this is the sample used in our analysis (except in the placebo tests). Because the nal pre-treatment wave of ENCELURB in 2004 is ve years prior to wave 1 of the debit card roll-out, we supplement our parallel trends test in ENCELURB with data for the intervening period ( ) from the Encuesta Nacional de Ingresos y Gastos de Hogares (National Household Income and Expenditure Survey; ENIGH), a repeated cross-section of between 20,000 and 30,000 households that is representative at the state level. We merge the publicly available ENIGH with restricted-access locality identiers provided by the Instituto Nacional de Estadística y Geografía (National Institute of Statistics and Geography; INEGI) to determine which 11 For each account in each bimester, this is the daily average balance, i.e. the end-of-day balance summed over all days in the bimester and divided over the number of days in the bimester. 12 This survey has been used by several papers including Angelucci and Attanasio (2013) and Behrman et al. (2012). 9

11 surveyed households were in treatment and control localities, and restrict the analysis to the poorest 20 percent of surveyed households to proxy for Oportunidades recipients. To test whether the delayed savings eect and increasing propensity to save over time can be explained by building trust or learning to use the technology, we use the Encuesta de Características Sociodemográcas de los Hogares Urbanos (Survey of Urban Households' Sociodemographic Characteristics; ENCASDU), a household survey conducted by Oportunidades at the end of 2010 which surveyed 8788 households across rural, semi-urban, and urban areas; of these, 1674 received Oportunidades benets in savings accounts tied to debit cards at the time of the survey. We also use the Encuesta Medios de Pago (Payment Method Survey), a survey conducted by Oportunidades in 2012 aimed at eliciting satisfaction with and use of the debit cards (or, in urban areas too far from a Banse branch or semi-urban and rural areas, the pre-paid cards not tied to a bank account). The Payment Methods Survey included 5381 households, drawn by stratied (by payment method and locality) random sampling from all Oportunidades beneciaries; of these, 1641 received their benets on debit cards tied to savings accounts, and Figure 3b shows the timing of the household survey data we use relative to the roll-out of debit cards. 4 Identication and estimation In order to identify causal impacts of the debit cards on savings in the Banse account, we use a period-by-period dierence-in-dierences (DID) strategy, taking advantage of the staggered expansion of debit cards. The identication assumption is that the beneciaries that got the card rst would have had the same savings and use of the account as those who got it later if it wasn't for receiving the card. The identication assumption is inherently untestable, but we follow standard practice and do two kinds of checks to ascertain its plausibility. The rst is the test in levels shown in Table 1 which compares means for treatment and control groups at the level of the locality and at the level of the individual account. As discussed above, most means do not have a statistically signicant dierence between wave 1, wave 2, and control localities; there is a dierence, however, in population, transfer amount, and percent of the transfer withdrawn. (For percent of the transfer withdrawn, the F-test of equality between the three means is rejected, and a test of equality of wave 1 and the control is rejected, but the test of equality between wave 2 and the control is not rejected.) The second testmore relevant for applying a DID identication strategyinvolves 10

12 showing that trends in savings and transactions were parallel before treatment for wave 1, wave 2, and the control. We perform these tests in Section 5 and show that pre-treatment trends are indeed parallel. The similarity of savings in the treatment and control groups before treatment contrast sharply with the dierence after the ATM card is awarded. The fact that results for two waves in dierent years are similar suggests this is not an artefact of a shock in a particular year. After showing the large increase in average balances, we control for transfers received and lagged balance to estimate the marginal propensity to save out of the transfer using a similar period-byperiod DID specication. The propensity to save is increasing over time with the card and there is a delayed eect in wave 1 before savings in wave 1 accounts are higher than in the control group, both of which suggest some kind of learning is taking place. If account holders were suddenly shifting all their cash under the mattress to the account, we would expect a once-and-for-all jump in the balance in the account. If they were only responding to a change in transaction costs by increasing their number of withdrawals as in the Baumol-Tobin model, we would see a once-and-for-all increase in the marginal propensity to save. Instead, we see rising savings for at least two years, as far as we can see in our data. We test and reject a supply side explanation for this gradual increase, according to which the increase in account savings results not from increasing trust but from an expansion of ATMs and branches (bank infrastructure) in treated localities. We rst estimate a DID model where we predict bank infrastructure (number of total ATMs, total bank branches, Banse ATMs, or Banse branches) with leads and lags of an indicator variable that equals 1 if the locality has switched to debit cards. We reject that ATM expansion was rst implemented in localities with more bank infrastructure (selection) or that the infrastructure grew more after the debit card in those localities. We also regress savings in our accounts against the expansion of bank infrastructure as an explanatory variable and do not nd a signicant relationship. Finally, we show that our results are unchanged when we control for the time-varying number of ATMs or bank branches on the right hand side of the regression. This means that the debit card itself played the main role, not new ATMs or branches. If the supply side is not generating the result, why then is the demand for formal savings increasing? We consider two possibilities. The rst is a pure transaction cost explanation, while the second involves an increased willingness to leave more money on the account. One simple 11

13 way to model the relationship between withdrawal costs and savings is to use the classic Baumol (1952) and Tobin (1956) model of demand for money. In that model, there is a benet to leaving money in the account (interest or avoiding theft, or perhaps avoiding temptation spending) but also a transaction cost of going to the ATM or bank branch to withdraw the money. The consumer optimally trades-o these two components. If the indirect transaction cost of withdrawing money decreases by having the ATM cardwhich is the case for most of the households in our study, who live closer to an ATM of any bank than a Banse branchthen the consumer optimally chooses to leave more money in the savings account and withdraw smaller amounts more frequently. So the average balance in the account increases mechanically not by withdrawing less pesos in the period, but purely by withdrawing more often. We call this the mechanical eect. The Baumol-Tobin model predicts that withdrawals will increase when consumers have an ATM card, a prediction that is borne out in the data: the average number of withdrawals per bimester increases from 1.02 at baseline to 1.53 with the card. Nevertheless, the majority of beneciaries continue making one withdrawal even with the card, and those who make two withdrawals still withdraw 71 percent of their transfer during the rst withdrawal (see the Appendix). Thus, we nd that quantitatively that this mechanical eect explains at most 10% of the increase in average balances, while the rest is due to decreases in the total amount withdrawn over the period (i.e., an actual increase in savings). The proportion of the transfer withdrawn over the course of the bimester falls from 98.9 percent to 85.1 percent (averaging over all bimesters in which an account holder has the card); in other words, the average proportion of the transfer saved increases from 1.1 percent to 14.9 percent. Are people learning how to use the technology, as Oportunidades program ocials speculated when we showed them our results from the administrative bank account data? Alternatively, are they learning to trust that their money will not disappear from the account if they save? We explore these two kinds of learning, which could generate the result of a delayed savings eect and increasing propensity to save over time. The rst we call operational learning; it involves the gradual understanding of how to use the ATM card, memorizing its personal identication number (PIN), etc. The second involves learning that risk of getting the money stolen in the form of hidden fees, operational errors, or malfeasance by the bank is lower than initially believed. These two explanations are observationally equivalent in the savings data, so that any hope of distinguishing 12

14 them has to bring additional data to bear. The rst data set we use is the ENCASDU, which includes survey questions that directly ask beneciaries if they save in their Banse accounts, and if not, why not. The data show clearly that less people report that they do not save because they do not trust the bank after they have had the card for at least 6 months, while the number of people reporting that they do not save due to a lack of knowledge or fear that the program will drop them for being not poor enough if they accumulate savings (another potential channel pointed out by Oportunidades program ocials) is both initially low and constant over time. Next, to explore mechanisms we test whether balance checks and actions that indicate knowledge of the technology change over time using the Payment Methods Survey, and nd that both the total number of balance checks and the number of trips to the ATM exclusively to check the account balance (without withdrawing any money) decrease after six months with the card. Importantly, we do not nd corresponding changes in proxies for knowledge like reporting that the ATM is hard to use, that they get help using the ATM, or that they know they can save in their account (although there is an increase in the proportion who know their PIN). We then test whether the self-reported balance check results are consistent with the administrative data on bank account transactions, and nd that individuals check their balance often when they get the cardconsistent with using the better monitoring technologybut gradually check it less and less. Operational learning would predict the opposite: all else constant, the easier it becomes is to check the balance as you learn, the more you should check it. Showing that savings in the account are radically higher as a result of having an ATM card and that trust is a likely driver of the eect is an important contribution to our understanding of the determinants of formal savings and the literature on trust and its eects on economic transactions. But one question that remains is if this increase of savings in the account constitutes an increase in total household savings or a shift from informal to formal savings. Total savings may rise because there is a new available asset in the portfolio which may have a good risk-return prole not available before and/or because it provides a commitment mechanism by reducing the amount of cash on hand. We test whether total savings increases using data from consumption, income and assets for Oportunidades beneciaries from the ENCELURB survey using a DID estimation strategy. Strikingly, our result for the increase in savings in the survey is very close quantitatively to the increase in savings in the account, which suggests that the increase in savings in the Banse 13

15 account is an increase in overall savings rather than a substitution from informal savings. Importantly, we document that the increased saving is not driven by higher income (which shouldn't be aected by the ATM) but by lower current consumption, and that the goods with a statistically signicant reduction in their consumption are the primary temptation goods discussed by Banerjee and Mullainathan (2010): alcohol, tobacco, and sugar. 5 Results for formal savings in the account 5.1 Eect on account balance and propensity to save Figure 5 presents the data of the time series of average balances; even the raw data is very telling. Panel (a) compares the rst wave of debit card recipients to the control group, while Panel (b) comapres the second wave to the control. Strikingly, average balances increase sharply for the rst wave after receiving the card, but the eect is not immediate: it begins three to four bimesters after receiving the card and the larger increase happens after a year with the card. By October of 2011, wave 1 has average balances of around 2000 pesos, over three times that of the control group. A similar pattern is present in wave 2, although we have information for less bimesters after the switch to debit cards. Although our data on average daily balances is by bimester, some payments get shifted to the end of the prior bimester, so we group adjacent bimesters into four-month periods for the remainder of the analysis. 13 Concretely, we estimate 9 y ijt = λ i + δ t + φ k T j I(k = t) + ε ijt (1) k=1 where y ijt is the average balance in account i from locality j over period t (specically, end of day balances were averaged over the number of days in the bimester by Banse, and we average the average balances over the two adjacent bimesters that make up the four-month period), λ i are account level xed eects which control for observable and unobservable time-invariant characteristics of the 13 If the bimesters are numbered consecutively with January-February 2009 as bimester 1, the payment shifts are always shifted from an odd-numbered bimester to the end of the preceding even-numbered bimester; since our data begin with an odd-numbered bimester, we thus create the rst period as January-February 2009, then each subsequent period as a four-month period: March-June 2009, July-October 2009, November 2009-February 2010, etc., so that payment shifts always occur within a four-month period. In our gures, the periods are labeled with just one of the months in the period to avoid clutter. 14

16 beneciaries, γ t are time-period dummies that control for general macro trends such as bimesterspecic shocks that aect both treatment and control groups, T j = 1 if locality j is a treatment locality, T j I(k = t) are time period dummies for treated localities to pick up the dierence in balances between treatment and control localities, and ε ijt are clustered by Banse branch. Since one time period dummy must be omitted from (1), we follow the standard procedure of omitting the four-month period immediately preceding the change to cards. We estimate (1) separately for wave 1 and wave 2. The coecients of interests are the φ k s, which measure the average dierence in balances between the control and treatment group in bimester k. The raw data clearly suggest that pre-treatment trends of savings were parallel across control and treatment groups before getting the card; we test this statistically by testing φ 1 = = φ l 1 = 0 where l is the period of switch. (In wave 1, l is the fourth four-month period, November 2009-February 2010, and in wave 2 it is the seventh fourmonth period, November 2010-February 2011.) Figure 6 plots the φ k s and shows that pretreatment coecients are, in most periods, not individually dierent from zero, and we cannot reject that pretrends are equal between treatment and control: the p-value for the F-test of φ 1 = = φ l 1 = 0 is for wave 1 and for wave 2. It also shows that the increase in savings is gradual and large, reaching 1,500 pesos of extra savings in the account after 2 years. The magnitudes are similar across the two waves (with the exception that we observe a quicker response to saving in the second wave), although we only have information one year after receiving the card for wave 2. To measure the propensity to save, we control for the amount received in transfers each period. This is important since there is a large amount of variation in transfers received within accounts over time, as well as between accounts (Figure 8a). The variation within an account over time can be explained by local elections in certain localities, 14 compliance with program conditions, 15 payment amounts varying depending on the time of year, 16 and dierences in family structure. One way to take into account dierent transfers is to estimate the propensity to save out of the transfer. That is, how much out of the transfer is saved and how does this propensity to save evolve over 14 When there is an election, Oportunidades has to give the transfer in advance, so that there is no payment close to the election month. In practice, this means that beneciaries receive no payment in the bimester of the election and an additional payment toward the end of the preceding bimester. 15 If a family does not comply with program conditions such as school attendance and health check-ups, the payment is suspended, but if the family returns to complying with the conditions, the missed payment is added into a future payment. 16 For example, the program includes a school component that is not paid during the summer, and a school supplies component that is only paid during one bimester out of the year. 15

17 time after treatment? In the spirit of asset accumulation models, we assume that savings in period t is a function of savings in period t 1 and the transfer in period t, and thus estimate y ijt = λ i + δ t + θy ij,t α k T j I(k = t) + γ k T ransfer ijt I(k = t) (2) k=1 k=1 9 ψ k T ransfer ijt T j I(k = t) + ɛ ijt k=1 where y ij,t 1 is the lagged balance and T ransfer ijt is the Oportunidades transfer amount deposited into account i in period t. The estimated marginal propensity to save out of transfer income is α k /µ k + ψ k, where µ k is average transfers in bimester k; Figure 7 plots the α k /µ k + ψ k estimates along with their condence intervals. Standard errors for the plotted point estimates are computed using the delta method. As before, we omit the four-month period immediately preceding the change to cards and estimate (2) separately for wave 1 and wave 2. In Figure 7, the propensity to save out of the transfer increases over time and increases substantially in two years. 5.2 The supply side as an explanation We have shown that beneciaries' savings are increasing in time after getting the debit card, and that they are leaving in the account a larger proportion of their transfer. We have interpreted this as a higher demand for savings in the account having to do with debit card expansion. In Section 6 we discuss dierent alternatives as to why the demand for savings might shift. Here we want to evaluate the possibility that the gradual increase in savings arises from a a gradual decrease in the cost of saving in the account. One possible explanation for the increase in savings over time is that banks gradually expanded complementary infrastructure in localities where treated beneciaries live. Depending on the costs of each branch and ATM machine, this could be a prot-maximizing response to the increase in the number of debit card holders in treated localities. The increasing time-prole of savings could be the result of the staggered expansion of this infrastructure, not increased trust. If this is so, then the increase in savings would have to be reinterpreted not only as the eect of debit cards 16

18 but of the expansion of the whole enabling technology. Using quarterly data for each municipality on the number of bank branches and ATMs for Banse and all other banks, we test if there was indeed a contemporaneous expansion of infrastructure and if this was correlated geographically with Oportunidades debit card expansion or with savings in our accounts. We rst test for a relationship between the roll-out of ATM cards and a supply-side expansion of banking infrastructure (ATMs and bank branches) 17 by estimating y jt = λ j + δ t + 4 k= 4 β k D j,t+k + ε jt, where y jt is the number of total ATMs, total bank branches, Banse ATMs, or Banse branches in municipality j in quarter t and D jt equals one if at least one locality in municipality j has Oportunidades debit cards in quarter t. We include one year (four quarters) of lags and one year of leads to test for a relationship between bank the debit card roll-out and bank infrastructure. For this test, we use data on the number of ATMs and bank branches by bank by municipality by quarter from the Comisión Nacional Bancaria y de Valores (CNBV), from the last quarter of 2008 through the last quarter of 2013 (since the roll-out continued in 2012, with what we refer to as control group localities receiving debit cards). We separately test whether lags of credit card receipt predict banking infrastructure (i.e., whether there is a supply-side response to the roll-out of debit cards) by testing β 4 = = β 1 = 0 and whether leads of credit card receipt predict banking infrastructure (i.e., whether debit cards were rst rolled out in municipalities with a recent expansion of banking infrastructure) by testing β 1 = = β 4 = 0. We nd evidence of neither relationship, failing to reject the null hypothesis of each test for each of the four dependent variables (Table 2). Another potential explanation is that together with the expansion of the accounts, Oportunidades changed other aspects of the program dierentially for those that got the debit cards. We already showed that the amounts of transfers did not change dierentially. After consulting with Oportunidades, we conrmed there were no other changes in the rules of the program. Oportunidades did incorporate more beneciaries during our sample period, but we study beneciaries that already were included in the program and received their benets through savings accounts prior to 17 We do not test an expansion of point of service (POS) payment terminals because the data on POS terminals by municipality does not begin until 2011, toward the end of our study period. 17

19 this expansion, and there is no reason there should be externalities to the existing beneciaries and less so externalities that dierentially impact treatment and control beneciaries. All in all, it seems that supply side changes do not explain why savings increase. 5.3 Deposits, withdrawals, and the mechanical eect Higher average balances, which we are using as a measure of savings, could arise from larger peso amounts of deposits, lower peso amounts of withdrawals, or from changes in the time prole of the two, without changing total peso amounts of either. We call the latter channel the mechanical eect as it can arise mechanically, purely from a reduction in the transaction cost of withdrawal. Motivated by the Baumol (1952) and Tobin (1956) model of demand for money and savings in the face of transaction costs, the debit card decreases the transaction cost of withdrawals, which implies more frequent withdrawals of lesser amounts and therefore a higher average amount in the account even if the total peso amount withdrawn per period remains constant (i.e., even when savings between one period and the next have not changed). Because we use average account balances (account balance at the end of the day, averaged over all days in the two-month pay period) as our measure of savings, our result could be driven by this mechanical eect. The card allows the beneciary to withdraw from the ATM of any bank, and many beneciaries live substantially closer to an ATM than to a Banse branch. This should reduce the cost of withdrawing money substantially. A simple comparison of the distribution of withdrawals in Figure 10 shows that the number of withdrawals is indeed larger for the treatment group: prior to the distribution of cards, 80 to 90 percent of beneciaries in both the treatment and control groups make only one withdrawal; after cards are distributed, the withdrawal patterns of the control group do not change, while the proportion of individuals in the treatment group withdrawing two or more times rises to about 30 percent. We now perform a back of the envelope calculation to measure the size of the mechanical eect. To calculate this number we do the following counterfactual exercise: we take as given the total peso amount withdrawn in the control group but impose the distribution of the frequency of withdrawals of the treated group, who change their distribution and frequency of withdrawals because of the debit card. Even after receiving the card, 72 percent of beneciaries continue to make just one withdrawal, while 22 percent make 2 withdrawals, 6 percent make 3 withdrawals, and 18

20 a negligible number of beneciaries make more than 3 withdrawals. For the 22 percent that make 2 withdrawals, on average they withdraw 71 percent of the total they withdraw over the bimester during the rst withdrawal and 29 percent during their second withdrawal, which is, on average, 9 days after the rst withdrawal. The 6 percent that make 3 withdrawals take 45 percent of the total amount they withdraw during the rst withdraw, which is on average 9 days before the second withdrawal when they extract 33 percent of the total they withdraw during the bimester, then after another 7 days they withdraw the remaining 22 percent of the amount they will withdraw during that bimester. This simple calculation implies that controls would have only 20 extra pesos in savings just from a higher withdrawal frequency; intuitively, because only 28 percent is switching to more than one withdrawal with the debit card, and because those that do switch still withdraw the bulk of their benets during the rst withdrawal and don't wait a full half or one-third of the bimester until making the next withdraw, the mechanical eect is not large. The eect amounts to just 10 percent of the extra savings we documented in Section 5 in the rst periods in which we see a positive eect of the debit cards on savings, and just 1 percent of the extra savings after two years. 6 Learning to use vs. learning to trust 6.1 Direct survey evidence Why do we see a delayed savings eect after receiving the debit card, and why does the marginal propensity to save out of the transfer gradually increase with time? We conjecture that learning is at play and explore two kinds of learning: operational learning and what we call learning to trust. The rst involves knowledge of how to use the debit card and ATM, memorizing the card's PIN, etc. The second involves learning that risk of getting the money stolen in the form of hidden fees, operational errors, or nefarious behavior by the bank is lower than initially believed. We nd evidence that beneciaries use the card to check their account balances, and that it thus provides them with a technology to monitor bank behavior, ensure that their money is not disappearing, and subsequently build trust in the bank. We rst use a survey that directly asks beneciaries Do you leave part of the monetary support from Oportunidades in your bank account? and, if the response is no: Why don't you keep part 19

21 of the monetary support from Oportunidades in your Banse bank account? 18 In the Payment Methods Survey, there are questions that allow us to explore the mechanisms behind learning to use the technology and learning to trust: What have been the main problems you have had with the ATM?; In general, does someone help you use the ATM?; Do you know your PIN by heart?; Did they tell you that with the card you have a Banse savings account?; In the last bimester, how many times did you check your balance?. Both surveys are cross sections and we compare answers to these questions as a function of time with the debit card, which is determined by the locality in which a beneciary lives. We estimate y i = α + γi(card 6 months) i + u i, where three regressions are run in which the dependent variable equals 1 if the beneciary reports not saving due to (i) a lack of knowledge, (ii) fear they will be dropped from the program, or (iii) lack of trust. We estimate the unconditional probability, i.e. beneciaries who report saving are included in the regression with y i = 0. Standard errors are clustered at the locality level. We test the null hypothesis γ 0, where a rejection of the null would imply that the dependent variable we are testing, which is related to either learning to use the technology or learning to trust the bank, changes over time with the card. Figure 11 and Panel (a) of Table 3 show the results. The rst thing to note is that low knowledge is rarely cited as a reason for not saving, while lack of trust is cited by 27 percent of those who do not save. Second, the proportion who report not saving due to a lack of knowledge does not change over time; in contrast, trust increases gradually with experience: beneciaries with more than 6 months with the card are 36 percent less likely to report not saving due to low trust than those with less than 6 months with the card. A third hypothesis that we test, which was suggested by Oportunidades ocials as a possibility, is that individuals fear that accumulating savings will make them ineligible (viewed as not poor enough) for the program. Here, we also nd that not many beneciaries report this as a reason for not saving in their Banse account, and there is no statistically signicant change in this proportion over time. 18 The second question includes pre-written responses and an open-ended response. An example of an answer coded as lack of trust is Because if I don't take out all the money I can lose what remains in the bank. An example of an answer coded as lack of knowledge is They didn't explain the process for saving. 20

22 Next, we explore mechanisms behind learning to trust vs. operational learning using the 2012 Payment Methods Survey. We use the same specication as above, but because this survey was conducted in 2012, those with the card for at least 6 months now include both wave 1 and wave 2, while beneciaries in some of the localities we treat as control localities throughout this paper make up the group with cards for less than 6 months. We use the same specication as above, with y i equal to: (i) the self-reported number of balance checks over the past bimester; (ii) the self-reported number of balance checks without withdrawing any money over the past bimester; and dummies if the respondent reports (iii) it is hard to use the ATM; (iv) she gets help using the ATM; (v) she knows her PIN; (vi) she knows she can save in the account. Both the number of balance checks and number of balance checks without withdrawing decrease over time with the card. Making trips to the ATM specically to check the account balance (i.e. making a balance check without withdrawing any money) decreases by 36 percent after six months compared to the rst 6 months, while most measures that indicate knowledge of how to use the technology do esnot change over time: the proportion who report it is hard to use the ATM, that they get help using the ATM, and that they know they can save in the account does not change, although there is a statistically signicant increase in the proportion who know their PINs. Taken together, these results suggest that trust in the account increases gradually after getting the ATM card, and that balance checks are used to monitor the bank and build trust in the account. 6.2 Administrative data on balance checks The self-reported balance checking results are consistent with those in the administrative bank account data. Figure 12 plots the number of times people check their balance as a function of time with the card; we observe that during each four month period after they receive the card, the number of balance checks decreases. The two types of learning make dierent predictions regarding this variable. Operational learning means that it is easierless costlyfor a beneciary to check her balance as she learns to use the technology (e.g., by memorizing her PIN or learning how to use the ATM). Therefore, if anything, we might expect her to check her balance more over time. On the contrary, learning to trust predicts that although at the start an individual would check her balance often to monitor her savings, she learns that her money is still in the account and updates downward her belief about the risk of 21

23 losing money. With simple Bayesian learning, balance checking has decreasing marginal benet and therefore she checks her balance less over time. The evidence from Figure 12 supports the theory of building trust rather than the theory of learning to use the technology. 6.3 Learning from time with the savings account We have argued that the card allows beneciaries to build trust in the bank by monitoring the bank's activity through balance checks. An alternative explanation for the observed delayed savings eect and increasing marginal propensity to save over time is that as they accumulate time with the savings account (rather than with the card), they learn the benets of saving. This possibility is unlikely for a number of reasons. First, both treatment and control accounts are accumulating time with their savings accounts simultaneously. Second, because the savings accounts were mainly rolled out between 2002 and 2004 (Figure 3), most beneciaries had already accumulated several years with the account by 2009, when our study begins. Indeed, the median date of account opening in our 342,709 accounts is October 18, 2004, and less than 5 percent of accounts had existed for less than two years when they received debit cards. Third, our results from Section 5 include account xed eects, so any time-invariant eect of having the account for a longer period of time would be absorbed. Fourth, to test for a time-varying eect of having the account for a longer period of time, we test whether results vary when we run the analysis separately for two groups: those who have had the account for more vs. less time. We use the median date of account opening to split the accounts into these two groups, and nd that results are very similar. Appendix Figure A1 shows the equivalent of Figure 7 separately for older accounts (panels (a) and (b)), opened before the median date of October 18, 2014, and younger accounts (panels (c) and (d)) opened on or after that date. 7 New savings vs. savings shifting The increase in formal Banse account savings might come at the expense of other types of savings that the household is already conducting, in such a way that total savings is not aected. This question is relevant not only if one is concerned with total household savings, but also to understand the mechanics through which the eect on formal savings is operating and as a rst step towards thinking about the broader welfare implications of providing a formal savings account with a debit 22

24 card. This section addresses two questions: rst, does the provision of the debit card and the increase in formal savings it causes increase total savings or is it merely a substitution from other forms of saving? Second, if it is total savings, where is the increase coming from? To address these questions, we use Oportunidades' ENCELURB panel survey, conducted in four waves during the years 2002, 2003, 2004 and November 2009 to February This survey is conducted by Oportunidades and has comprehensive modules on consumption, income, and assets for about 6000 households in urban and semi-urban areas. 19 As before, we use a dierences-in-dierences strategy where we seek to explain changes consumption, savings, and income across beneciaries, exploiting the dierential timing of debit card receipt. Because the ENCELURB was conducted after wave 1 localities had received cards but before wave 2 or control localities had received cards, we compare those with cards (wave 1) to those who have not yet received cards (waves 2 and control), referring to them as treatment and control, respectively, in this section of the paper. The identication assumption is that in the absence of the debit card, treatment and control groups would have experienced similar changes in consumption, income, and assets. We show that trends of consumption are parallel before the intervention in Figure 13; panel (a) shows that, prior to treatment, there are parallel trends in consumption in treatment vs. control localities. Because a number of years pass between our last pre-treatment wave (2004) and the intervention (2009), we supplement the parallel trends test with data from ENIGH in panel (b) for the intervening period from We cannot statistically reject that trends are parallel. The advantage of also using the ENIGH is not only that it has a 30 page consumption module and is the ocial way to measure expenditure in Mexico, but also that we have information for the years leading up to the treatment. Appendix Figure!pending! shows that trends are also parallel for income and savings. Having shown parallel trends, we now test whether there was an increase in overall savings. We nd that there was an increase in overall savings, and that while income followed a similar trajectory in the treatment and control groups, the treatment group reduced its consumption. We estimate 19 The 2002, 2003, and 2004 waves had around 17,000 households, but due to budget constraints the number of localities was cut for the wave. The consumption, income, and assets modules of a similar survey conducted in rural areas have been used by Angelucci and De Giorgi (2009), Attanasio et al. (2013), Gertler et al. (2012), and Hoddinott and Skouas (2004), while these modules from the ENCELURB have been used by Angelucci and Attanasio (2013) and Behrman et al. (2012). 23

25 y ijt = φ i + δ t + γd jt + ν ijt (3) separately for ve dependent variables: consumption, income, savings (constructed as income minus consumption), purchase of durables, and an asset index. All variables except assets are measured in pesos per month. Durables correspond to the rst principal component of dummy variables indicating ownership of the assets that are included in all rounds of the survey questionnaire: car, truck, motorcycle, TV, video or DVD player, radio, washer, gas stove and refrigerator. The specication includes household and year xed eects, and standard errors are clustered at the locality level. If the increase in formal savings constitutes an increase in total savings then we expect γ > 0 for total savings (dened as income minus consumption), and if we observe γ = 0 for income we expect γ < 0 for consumption. If there is no substitution of savings from assets (and if they are not using the formal savings accounts to save up for assets, at least in the short run), we expect γ = 0 for the purchase of durables (which measures a ow) and the asset index (which measures a stock). This is indeed what we nd. Figure 14 shows that consumption decreased almost 200 pesos on average and is statistically signicant at the 5% level. Meanwhile, there is no eect on income; we also test the dierence in the coecients of consumption and income using a stacked regression; although both are noisily measured, the dierence is signicant at the 10% level. Purchase of durables and the stock of assets do not change, ruling out a crowding out of these forms of saving. The increase in savings, measured as income minus consumption, is estimated at slightly more than 200 pesos, and is signicant at the 5 percent level. The results in Figure 14 are from our preferred specication where we winsorize the dependent variable at 5 percent (specically, at the 95th percentile, as well as the 5th percentile if the variable does not have a lower bound of 0) to avoid letting our results be driven by outliers, and where we interact household characteristics with a linear time trend. The eects are robust to using the raw data or winsorizing 1% or 5% of the sample (we follow Kast and Pomeranz (2014) who show the robustness of results to these three possibilities for their savings measures)with or without the baseline characteristics interacted with a linear time trend that we include in the preferred specifcation, as well as including separate linear time trends for each municipality The household characteristics interacted with a linear time trend in this robustness check are measured at baseline 24

26 These results mean that total savingsnot just formal savingsincrease, and that this increase in being funded by lower consumption and not by selling durables. A back of the envelope calculation reveals that the magnitude of the increase in monthly savings from this household survey is in line with the average increase of savings in the account from the administrative data: from the propensity to save specication, after 1 year, beneciaries who have received cards save 25.1 percent of their transfer more than the control group. Using ENCELURB, transfers are, on average, 20.2 percent of income for the treatment group, implying that the savings eect in the Banse administrative data is about 5.1% of income. The eect for savings (income minus consumption) in the ENCELURB household survey data shown in Figure 13 equates to 4.8 percent of income. Taken at face value, this suggests that most of the increase in savings in the account is new saving. This result is consistent with other studies where formal savings products were oered, which found that the increased savings in these products did not crowd out other forms of savings (Ashraf et al., 2006, 2010; Dupas and Robinson, 2013a; Prina, 2015). We run two placebo tests by estimating regression (3) for dierent samples from the same household survey: using poor non-oportunidades households in the treated vs. control localities, and also by using the award of a pre-paid (not debit) card in other urban localities (Figure 15). We nd no eects for these placebos. The rst placebo shows that it is not locality-specic shocks driving our results, as these locality-specic shocks would be expected to aect poor non-oportunidades beneciaries as well. The second placebo shows that the result is not driven by an informational/salience eect, where the card serves as a salient reminder to the beneciary of her savings goals. Instead, the dierence between a debit and a pre-paid card is that beneciaries cannot save with a pre-paid card, only withdraw the full transfer amount. It seems therefore that it is the increased convenience of accessing savings at any ATM and the ability to monitor the bank and build trust which causes the dierence. Finally, because the accounts pay no interest, but there was clearly an unmet demand for savings among program beneciaries, we explore why they were not able to save before (for example, under the mattress). If indeed the provision of a debit card savings induces higher total savings through and include characteristics of the household head (whether the household head worked, a quadratic polynomial in years of schooling, and a quadratic polynomial in age), whether the household has a bank account, variables used to measure poverty by Oportunidades (the proportion of household members with health insurance, the proportion aged 15 or older that are illiterate, the proportion aged 6 to 14 that do not atted school, the proportion aged 15 or older with incomplete primary education, and the proportion aged 15 to 29 with less than 9 years of schooling), and dwelling characteristics (dirt oor, no bathroom, no water, no sewage, number of occupants per room). 25

27 decreased consumption, we would expect that it inuences dierent components of consumption dierentially. We dene temptation goods as alcohol, tobacco, and sugar, which are the most frequently cited temptation goods in Banerjee and Mullainathan (2010), and we estimate whether the proportion of income spent on these categories decreases. 21 We estimate a DID specication with household and year xed eects and standard errors clustered at the locality level, as before. Consumption gijt Income ijt = λ gi + δ gt + γ g D jt + ν gijt (4) We nd that consumption of temptation goods (and transportation) decrease more than the other goods in the treatment group relative to the control (Figure 16), and that these are the only two categories where the decrease in consumption is statistically signicant. We interpret this result as evidence that it is dicult to save informally due to self-control problems, and that these problems can be partially solved by access to a formal savings account (but that low indirect transaction costs and trust in the bank are necessary conditions for these formal savings accounts to be usable). This nding is consistent with the demand for commitment savings devices (Ashraf et al., 2006; Bryan et al., 2010, e.g.,) and the nding that microcredit decreases temptation good consumption (Angelucci et al., 2015; Augsburg et al., 2015; Banerjee et al., 2015). 8 Conclusion Although trust in nancial institutions is by no means a sucient condition to enable the poor to save, our ndings suggest that it is a necessary condition. A lack of trust in banks could explain why a number of studies have found modest eects of oering savings accounts to the poor, even when these accounts have no fees or minimum balance requirements. Debit cards, a simple technology with high scale-up potential, provided beneciaries of Mexico's large-scale cash transfer program Oportunidades with a mechanism to monitor banks by checking their balances at any bank's ATM; once beneciaries built trust in banks, they began to save and their marginal propensity to save increased over time. We nd that the observed increase in formal savings represents an increase in overall savings rather than a substitution from other forms of saving, and that beneciaries 21 Using proportions reduces noise and variance inherent in variables dened in pesos and therefore increases statistical precision. 26

28 reduce consumption of temptation goods, suggesting that saving informally is dicult and the use of nancial institutions to save helps solve self-control problems. The size of the savings eect, at 5% of income after one year with the debit card and 10% after two years, is larger than that of studies on various savings interventions such as subsidizing bank fees, increasing interest rates, and providing commitment savings devices. As a result, interventions that enable account holders to monitor banks and increase their trust in nancial institutions may be a promising avenue to enable the poor to save in the formal nancial sector. Debit cards and other forms of mobile money, which are simple, scalable technologies that are gaining traction in many developing countries, could thus be a highly eective means of increasing nancial inclusion among millions of government cash transfer recipients worldwide. 27

29 Figure 1: Low Trust in Banks Around the World Percent with low trust (60,100] (50,60] (40,50] (28,40] [0,28] No data Source: World Values Survey, Wave 6 ( ). Notes: N = 82,587 individuals in 60 countries. Low trust in banks is dened as not very much condence or none at all for the item banks in response to the following question: I am going to name a number of organizations. For each one, could you tell me how much condence you have in them: is it a great deal of condence, quite a lot of condence, not very much condence or none at all? Countries are divided into quintiles, with quintile cut-os rounded to the nearest percentage point in the legend. Darker shades indicate countries with a higher percent of the population reporting low trust in banks. No formal education Figure 2: Low Trust in Banks by Education Level in Mexico Incomplete Primary Complete Primary Complete Secondary Complete University Percent with low trust in banks Source: World Values Survey, Mexico, Wave 6 (2012). Notes: N = 1993 individuals. Low trust in banks is dened as not very much condence or none at all for the item banks in response to the following question: I am going to name a number of organizations. For each one, could you tell me how much condence you have in them: is it a great deal of condence, quite a lot of condence, not very much condence or none at all? 28

30 Figure 3: Timing of Roll-out and Data (a) Administrative Bank Account Data Oportunidades bank accounts without cards Oportunidades bank accounts with cards Bansefi account balances and transactions 1,000, , , , ,000 0 Jan 02 May 02 Sep 02 Jan 03 May 03 Sep 03 Jan 04 May 04 Sep 04 Jan 05 May 05 Sep 05 Jan 06 May 06 Sep 06 Jan 07 May 07 Sep 07 Jan 08 May 08 Sep 08 Jan 09 May 09 Sep 09 Jan 10 May 10 Sep 10 Jan 11 May 11 Sep 11 Jan 12 (b) Household Survey Data Oportunidades bank accounts without cards Oportunidades bank accounts with cards Payment Methods Survey ENCELURB ENIGH ENCASDU 1,000, , , , ,000 0 Jan 02 May 02 Sep 02 Jan 03 May 03 Sep 03 Jan 04 May 04 Sep 04 Jan 05 May 05 Sep 05 Jan 06 May 06 Sep 06 Jan 07 May 07 Sep 07 Jan 08 May 08 Sep 08 Jan 09 May 09 Sep 09 Jan 10 May 10 Sep 10 Jan 11 May 11 Sep 11 Jan 12 Source: Number of Oportunidades bank accounts with cards and without cards by bimester is from administrative data provided by Oportunidades. 29

31 Figure 4: Geographic Coverage and Expansion of Debit Cards Sources: Administrative data from Oportunidades on timing of debit card receipt by locality and shape les from INEGI. Notes: N = 275 localities (44 in control, 143 in wave 1, 88 in wave 2). The area of each urban locality included in the study is shaded according to its wave of treatment. Urban localities that were not included in the Oportunidades program at baseline or were included in the program but did not pay beneciaries through Banse savings accounts are not included in the gure or in our study. Figure 5: Evolution of Average Balances (a) Wave 1 vs. Control (b) Wave 2 vs Control 2000 Wave 1 Control 2000 Wave 2 Control Jan 09 Mar 09 May 09 Jul 09 Sep 09 Nov 09 Jan 10 Mar 10 May 10 Jul 10 Sep 10 Nov 10 Jan 11 Mar 11 May 11 Jul 11 Sep 11 Sources: Administrative data from Banse on average account balances by bimester and timing of card receipt. Notes: N = 5,834,468 account-bimester observations from 343,204 accounts. Average balances are winsorized at the 95th percentile. 0 Jan 09 Mar 09 May 09 Jul 09 Sep 09 Nov 09 Jan 10 Mar 10 May 10 Jul 10 Sep 10 Nov 10 Jan 11 Mar 11 May 11 Jul 11 Sep 11 30

32 Figure 6: Dierence between Treatment and Control in Average Balances 1500 (a) Wave 1 vs. Control 1500 (b) Wave 2 vs. Control Jan 09 May 09 Sep 09 Jan 10 May 10 Sep 10 Jan 11 May 11 Sep 11 Jan 09 May 09 Sep 09 Jan 10 May 10 Sep 10 Jan 11 May 11 Sep 11 Sources: Administrative data from Banse on average account balances by bimester and timing of card receipt. Notes: (a) N = 1,542,969 from 171,441 accounts. (b) N = 2,430,414 from 270,046 accounts. The gure plots φ k from y ijt = λ i + δ t + 9 φ k T j I(k = t) + ɛ ijt. k=1 Average balance over each four-month period is the dependent variable, and is winsorized at the 95th percentile. Whiskers denote 95 percent condence intervals. Black lled in circles indicate results that are signicant at the 5 percent level, gray lled in circles at the 10 percent level, and hollow circles indicate results that are statistically insignicant from 0. The period prior to receiving the card is the omitted period, which is why its point estimate is 0 with no condence interval. 31

33 Figure 7: Dierence between Treatment and Control in Marginal Propensity to Save Out of Transfer 0.6 (a) Wave 1 vs. Control 0.6 (b) Wave 2 vs. Control May 09 Sep 09 Jan 10 May 10 Sep 10 Jan 11 May 11 Sep 11 May 09 Sep 09 Jan 10 May 10 Sep 10 Jan 11 May 11 Sep 11 Sources: Administrative data from Banse on average account balances by bimester, transfer payments, and timing of card receipt. Notes: (a) N = 1,542,969 from 171,441 accounts. (b) N = 2,430,414 from 270,046 accounts. The gure plots α k µ k + ψ k from y ijt = λ i + δ t + θy ij,t α k T j I(k = t) + k=1 9 γ k T ransfer ijt I(k = t) k=1 9 ψ k T ransfer ijt T j I(k = t) + ɛ ijt k=1 where µ k is average transfers in period k. Average balance over each four-month period is the dependent variable, and is winsorized at the 95th percentile; the regression controls for lagged average balance and estimates eects as a proportion of the transfer amount received, where transfer amounts are also winsorized at the 95th percentile. Standard errors for the point estimates of α k µ k + ψ k are computed using the delta method. Whiskers denote 95 percent condence intervals. Black lled in circles indicate results that are signicant at the 5 percent level, gray lled in circles at the 10 percent level, and hollow circles indicate results that are statistically insignicant from 0. The period prior to receiving the card is the omitted period, which is why its point estimate is 0 with no condence interval. 32

34 Figure 8: Evolution of Deposit and Withdrawal Amounts (a) Amount Deposited by bimester (a) Wave 1 vs. Control (b) Wave 2 vs. Control 3500 Wave 1 Control 3500 Wave 2 Control Jan 09 Mar 09 May 09 Jul 09 Sep 09 Nov 09 Jan 10 Mar 10 May 10 Jul 10 Sep 10 Nov 10 Jan 11 Mar 11 May 11 Jul 11 Sep 11 Jan 09 Mar 09 May 09 Jul 09 Sep 09 Nov 09 Jan 10 (b) Amount Withdrawn by bimester 0 Mar 10 May 10 Jul 10 Sep 10 Nov 10 Jan 11 Mar 11 May 11 Jul 11 Sep 11 (a) Wave 1 vs. Control (b) Wave 2 vs. Control 3500 Wave 1 Control 3500 Wave 2 Control Jan 09 Mar 09 May 09 Jul 09 Sep 09 Nov 09 Jan 10 Mar 10 May 10 Jul 10 Sep 10 Nov 10 Jan 11 Mar 11 May 11 Jul 11 Sep 11 0 Jan 09 Mar 09 May 09 Jul 09 Sep 09 Nov 09 Jan 10 Mar 10 May 10 Jul 10 Sep 10 Nov 10 Jan 11 Mar 11 May 11 Jul 11 Sep 11 Sources: Administrative data from Banse on transactions by bimester and timing of card receipt. Notes: N = 5,834,468 account-bimester observations from 343,204 accounts. 33

35 Figure 9: Evolution of Deposit and Withdrawal Frequencies (a) Deposit frequency by bimester 2.5 Wave 1 Control (a) Wave 1 vs. Control 2.5 Wave 2 Control (b) Wave 2 vs. Control Jan 09 Mar 09 May 09 Jul 09 Sep 09 Nov 09 Jan 10 Mar 10 May 10 Jul 10 Sep 10 Nov 10 Jan 11 Mar 11 May 11 Jul 11 Sep 11 Jan 09 Mar 09 May 09 Jul 09 Sep 09 Nov 09 Jan 10 Mar 10 (b) Withdrawal frequency by bimester 0 May 10 Jul 10 Sep 10 Nov 10 Jan 11 Mar 11 May 11 Jul 11 Sep Wave 1 Control (a) Wave 1 vs. Control 2.5 Wave 2 Control (b) Wave 2 vs. Control Jan 09 Mar 09 May 09 Jul 09 Sep 09 Nov 09 Jan 10 Mar 10 May 10 Jul 10 Sep 10 Nov 10 Jan 11 Mar 11 May 11 Jul 11 Sep 11 0 Jan 09 Mar 09 May 09 Jul 09 Sep 09 Nov 09 Jan 10 Mar 10 May 10 Jul 10 Sep 10 Nov 10 Jan 11 Mar 11 May 11 Jul 11 Sep 11 Sources: Administrative data from Banse on transactions by quarter and timing of card receipt. Notes: N = 2,917,234 account-quarter observations from 343,204 accounts. 34

36 Figure 10: Distribution of Withdrawals frequency frequency Distribution of withdrawals BEFORE switch Wave 1 Control or more Distribution of withdrawals BEFORE switch Wave 2 Control (a) Wave 1 vs. Control (b) Wave 2 vs. Control Distribution of withdrawals AFTER switch or more Distribution of withdrawals AFTER switch or more or more Sources: Administrative data from Banse on transactions by bimester and timing of card receipt. Notes: N = 5,834,468 account-bimester observations from 343,204 accounts. 35

37 Figure 11: Trust and Knowledge Over Time with the ATM Card (a) ENCASDU (2010).4 Does not save in Bansefi account due to Debit card < 6 months ** Debit card > 6 months Lack of knowledge Fear of ineligibility Lack of trust (b) Payment Methods Survey (2012) 1.5 Building trust ** **.6 Debit card < 6 months Debit card > 6 months Learning the technology ** Balance checks Checks without withdrawing 0 Hard to use ATM Gets help using ATM Knows PIN Knows can save in account Sources: ENCASDU 2010 and Payment Methods Survey Notes: (a) N = (b) N = 1617, or less in some regressions if there were respondents who reported don't know or refused to respond (see Table 3 for number of observations in each regression). Balance checks are measured over the past bimester. Whiskers denote 95 percent condence intervals. Bars for debit card < 6 months are colored light blue in (a) because at the time of ENCASDU 2010, those with the card 6 months or less were in wave 2 localities; bars for debit card < 6 months are colored orange in (b) because at the time of Payment Methods Survey 2012, 36 those with the card 6 months or less were in control localities.

38 Figure 12: Balance Checks (Administrative Data) 2.5 Wave Wave Nov 09 Mar 10 Jul 10 Nov 10 Mar 11 Jul 11 Nov 11 0 Nov 09 Mar 10 Jul 10 Nov 10 Mar 11 Jul 11 Nov 11 Source: Aministrative transactions data from Banse. Notes: Number of balance checks per account tied to a debit card. Prior to receiving the card it was possible to check balances at Banse branches only, and balance checks at Banse branches are not recorded in our transactions data because they are free of charge. Figure 13: Parallel Pre-Treatment Trends in Household Survey Data (a) ENCELURB: Pre Treatment Consumption (pesos/week) Control Treatment (b) ENIGH: Pre Treatment Consumption (pesos/week) Control Treatment Year Year 37

39 Figure 14: Eect of the debit card on consumption, income, total savings, purchase of durables, and assets Difference in differences estimates Consumption * Income Savings = Income Consumption Purchase of Durables Pesos per month space Asset Index Standard deviations Sources: ENCELURB panel survey combined with administrative data on timing of card receipt and transfer payment histories for each surveyed beneciary household. Notes: N = 11, 243 (number of households = 2938). Dependant variables are measured in pesos per month, with the exception of the asset index. Asset index is the rst principal component of assets that are included in both the early (2002, 2003, 2004) and post-treatment ( ) versions of the survey: car, truck, motorcycle, television, video or DVD player, radio or stereo, washer, gas stove, and refrigerator. Whiskers denote 95 percent condence intervals. Black lled in circles indicate results that are signicant at the 5 percent level, gray lled in circles at the 10 percent level, and hollow circles indicate results that are statistically insignicant from 0. The * linking consumption and income denotes that a test of equal coecients from the consumption and income regressions is rejected at the 10 percent level using a stacked regression. Results are from the preferred specication of winsorizing variables at the 95th percentile (and 5th percentile for variables that do not have a lower bound of 0) and controling for baseline household characteristics interacted with a linear time trend. Raw results, winsorized at 1 percent, winsorized at 5 percent (without household characteristics interacted with a linear time trend), and winsorized at 5 percent with municipality xed eects interacted with a linear time trend are similar and available in Appendix Table A. All regressions include household and time xed eects, and standard errors are clustered at the locality level. 38

40 Figure 15: Placebo Tests (a) Placebo 1: Poor Non Beneficiaries Consumption Income Savings = Income Consumption Pesos per month (b) Placebo 2: Pre Paid Cards Consumption Income Savings = Income Consumption Pesos per month (c) Original Estimates (For Comparison) Consumption Income Savings = Income Consumption Pesos per month Sources: ENCELURB panel survey combined with administrative data on timing of card receipt and transfer payment histories for each surveyed beneciary household. Notes: (a) N = 1415 (number of households = 382); (b) N = 8862 (number of households = 2300); (c) N = 11, 243 (number of households = 2938). Whiskers denote 95 percent condence intervals. Black lled in circles indicate results that are signicant at the 5 percent level, gray lled in circles at the 10 percent level, and hollow circles indicate results that are statistically insignicant from 0. Results are from the preferred specication of winsorizing variables at the 95th percentile (and 5th percentile for variables that do not have a lower bound of 0) and controling for baseline household characteristics interacted with a linear time trend. Raw results, winsorized at 1 percent, winsorized at 5 percent (without household characteristics interacted with a linear time trend), and winsorized at 5 percent with municipality xed eects interacted with a linear time trend are similar and available in Appendix Table B. All regressions include household and time xed eects, and standard errors are clustered at the locality level. 39

41 Figure 16: Eect of the debit card on dierent categories of consumption Percent change in proportion of income spent on... Alcohol, tobacco, sugar Junk food, fats, soda Meat, dairy, produce Tortillas and cereals Entertainment Transportation Health and education Percent 40

42 Table 1: Comparison of Baseline Means Variable Control Wave 1 Wave 2 Di. Di. F-test W1C W2C p-value Panel A: Locality-level data Log population (0.11) (0.10) (0.16) (0.14) (0.19) Banse branches/100, = (0.28) (0.13) (0.23) (0.30) (0.36) % HHs in poverty =2.73 = (1.67) (0.75) (1.09) (1.82) (1.99) Occupants per room =0.07 = (0.04) (0.01) (0.02) (0.04) (0.04) Number of localities Panel B: Administrative bank account data Average balance (12.46) (56.24) (21.26) (55.33) (23.95) Number of deposits =0.02 = (0.01) (0.04) (0.03) (0.04) (0.03) Size of transfer (12.73) (20.16) (17.47) (23.67) (21.15) Number of withdrawals =0.01 = (0.01) (0.03) (0.02) (0.03) (0.02) Percent withdrawn = (0.18) (0.45) (0.71) (0.46) (0.72) Years with account = (0.08) (0.15) (0.25) (0.17) (0.26) Number of accounts 97,922 73, ,717 Sources: Census (2005), Banse branch locations (2008), poverty estimates from Oportunidades (based on 2005 Census), timing of card receipt by locality from Oportunidades, and administrative data from Banse. Notes: W1 = wave 1, W2 = wave 2, C = control, Di. = dierence. For the administrative data from Banse, baseline is dened as January 2009 to October 2009 (prior to any accounts receiving cards in the data from Banse). 41

43 Table 2: Supply-Side Response Total Banse ATMs Branches ATMs Branches Current quarter = =0.01 =0.01 (4.14) (0.30) (0.01) (0.02) 1 quarter lag = (4.11) (0.34) (0.01) (0.02) 2 quarter lag = (5.64) (0.36) (0.03) (0.01) 3 quarter lag = = (2.98) (0.26) (0.02) (0.02) 4 quarter lag = =0.03 (5.97) (0.50) (0.01) (0.03) 1 quarter lead =1.10 =0.12 = (3.66) (0.36) (0.00) (0.02) 2 quarter lead = (4.90) (0.34) (0.02) (0.01) 3 quarter lead = =0.03 =0.01 (8.00) (0.65) (0.01) (0.03) 4 quarter lead =0.01 =0.06 (10.32) (0.94) (0.03) (0.05) Mean control group F-test of lags [p-value] [0.29] [0.94] [0.61] [0.43] F-test of leads [p-value] [0.60] [0.78] [0.53] [0.62] Municipality xed eects Yes Yes Yes Yes Quarter xed eects Yes Yes Yes Yes Notes: indicates p < 0.1, p < 0.05, and p < The table shows β k from y jt = λ j + δ t + 4 k= 4 β k D j,t+k + ε jt where y jt is the number of ATMs or bank branches of any bank or of Banse in municipality j during quarter t, D jt = 1 if municipality j has at least one locality with Oportunidades debit cards in quarter t. The F-test of lags tests β 4 = = β 1 = 0; the F-test of leads tests β 1 = = β 4 = 0. 42

44 Table 3: Trust and Knowledge Over Time with the ATM Card Mean Has card N 6 months Panel A: ENCASDU Survey (2010): Doesn't save in Banse due to... Lack of knowledge = ,674 (0.010) (0.010) Fear of ineligibility ,674 (0.014) (0.015) Lack of trust ,674 (0.012) (0.044) Panel B: Payment Methods Survey (2012) Lack of trust Times checked balance ,493 (0.039) (0.105) Times checked balance without withdrawing ,490 (0.035) (0.093) Lack of knowledge Hard to use ATM ,617 (0.013) (0.025) Gets help using ATM ,612 (0.023) (0.048) Knows PIN = ,609 (0.017) (0.034) Knows can save in account = ,617 (0.023) (0.046) Notes: indicates p < 0.1, p < 0.05, and p < Standard errors are clustered at the locality level. The Mean column shows the mean for those who have had the card for more than six months; the Has card 6 months column shows the regression coecient on a dummy for those who have had the debit card for six months or fewer (i.e., the dierence relative to the mean column). The precise questions on trust and knowledge are as follows. In the ENCASDU, the questions are Do you leave part of the monetary support from Oportunidades in your bank account? and, if the response is no, Why don't you keep part of the monetary support from Oportunidades in your Banse bank account? The regressions presented here are not conditional on saving; those who report yes to the rst question are coded with trust and knowledge dependent variables of 0 and included in the regressions. The second question includes pre-written responses and an open-ended response (other; specify; 4% of the sample in this table responded using the open-ended option); both pre-written and open-ended responses were coded as lack of knowledge, fear of ineligibility, lack of trust, or another explanation for not saving. An example of an answer coded as lack of knowledge is They didn't explain the process for saving. An example of an answer coded as fear of ineligibility is Because if I save in that account they can remove me from the Oportunidades program. An example of an answer coded as lack of trust is Because if I don't take out all the money, I can lose what remains in the bank. In the Payment Methods Survey, each regression comes from a dierent survey question. These questions are: (1) Times checked balance: In the last bimester, how many times did you consult your balance? (2) Times checked balance without withdrawing: created by subtracting In the last bimester, how many times did you withdraw money from the ATM? from (1); (3) Hard to use ATM: responded The ATM is dicult to use (pre-written response) or a similar open-ended response to the question What have been the main problems you have had with the ATM? [Wait for a response and record up to three of the options]; (4) Gets help using ATM: In general, does someone help you use the ATM?; (5) Knows PIN: Do you know your PIN by heart?; (6) Knows can save in account: Did they tell you that with the card you have a Banse savings account? 43

45 References Angelucci, M., Attanasio, O., The demand for food of poor urban mexican households: Understanding policy impacts using structural models. American Economic Journal: Economic Policy 5, Angelucci, M., De Giorgi, G., Indirect eects of an aid program: How do cash transfers aect ineligibles' consumption? American Economic Review 99, Angelucci, M., Karlan, D.S., Zinman, J., Microcredit impacts: Evidence from a randomized microcredit program placement experiment by Compartamos Banco. American Economic Journal: Applied Economics 7, Arrow, K., Gifts and exchanges. Philosophy & Public Aairs 1, Ashraf, N., Spousal control and intra-household decision making: An experimental study in the Philippines. American Economic Review 99, Ashraf, N., Karlan, D., Yin, W., Tying Odysseus to the mast: Evidence from a commitment savings product in the Philippines. Quarterly Journal of Economics 121, Ashraf, N., Karlan, D., Yin, W., Female empowerment: Impact of a commitment savings product in the philippines. World Development 38, Attanasio, O., Di Maro, V., Lechene, V., Phillips, D., Welfare consequences of food prices increases: Evidence from rural Mexico. Journal of Development Economics 104, Augsburg, B., de Haas, R., Harmgart, H., Meghir, C., The impacts of microcredit: Evidence from Bosnia and Herzegovina. American Economic Journal: Applied Economics 7, Baland, J.M., Guirkinger, C., Mali, C., Pretending to be poor: Borrowing to escape forced solidarity in Cameroon. Economic Development and Cultural Change 60, 116. Banerjee, A., Mullainathan, S., The shape of temptation: Implications for the economic lives of the poor. NBER Working Paper URL: Banerjee, A.V., Microcredit under the microscope: What have we learned in the past two decades, and what do we need to know? Annual Review of Economics 5, Banerjee, A.V., Duo, E., Glennester, R., Kinnan, C., The miracle of micronance? Evidence from a randomized evaluation. American Economic Journal: Applied Economics 7, Barber, S.L., Gertler, P.J., The impact of Mexico's conditional cash transfer programme, Oportunidades, on birthweight. Tropical Medicine and International Health 13, Barham, T., Rowberry, J., Living longer: The eect of the Mexican conditional cash transfer program on elderly mortality. Journal of Development Economics 105, Baumol, W.J., The transactions demand for cash: An inventory theoretic approach. Quarterly Journal of Economics 66, Behrman, J.R., Gallardo-García, J., Parker, S.W., Todd, P.E., Vélez-Grajales, V., Are conditional cash transfers eective in urban areas? Evidence from Mexico. Education Economics 20, Behrman, J.R., Parker, S.W., Todd, P.E., Do conditional cash transfers for schooling generate lasting benets? A ve-year followup of PROGRESA/Oportunidades. Journal of Human Resources 46, Brune, L., Giné, X., Goldberg, J., Yang, D., forthcoming. Facilitating savings for agriculture: Field experimental evidence from malawi. Economic Development and Cultural Change. Bryan, G., Karlan, D., Nelson, S., Commitment devices. Journal of Development Economics 2, Chaia, A., Dalal, A., Goland, T., Gonzalez, M.J., Morduch, J., Schi, R., Half of the World is Unbanked. McKinsey & Company, New York. Dupas, P., Green, S., Keats, A., Robinson, J., forthcoming. Challenges in banking the rural poor: 44

46 Evidence from Kenya's Western Province, in: Edwards, S., Johnson, S., Weil, D.N. (Eds.), African Successes: Modernization and Development. University of Chicago Press, Chicago. Dupas, P., Robinson, J., 2013a. Savings constraints and microenterprise development: Evidence from a eld experiment in Kenya. American Economic Journal: Applied Economics 5, Dupas, P., Robinson, J., 2013b. Why don't the poor save more? Evidence from health savings experiments. American Economic Review 103, Fafchamps, M., Udry, C., Czukas, K., Drought and saving in West Africa: are livestock a buer stock? Journal of Development Economics 55, Fernald, L.C.H., Gertler, P.J., Neufeld, L.M., year eect of Oportunidades, Mexico's conditional cash transfer programme, on child growth, cognition, language, and behaviour: A longitudinal follow-up study. The Lancet 374, Gertler, P.J., Do conditional cash transfers improve child health? Evidence from PRO- GRESA's control randomized experiment. American Economic Review Papers and Proceedings 94, Gertler, P.J., Martinez, S.W., Rubio-Codina, M., Investing cash transfers to raise long-term living standards. American Economic Journal: Applied Economics 4, Guiso, L., Sapienza, P., Zingales, L., The role of social capital in nancial development. American Economic Review 49, Hoddinott, J., Skouas, E., The impact of PROGRESA on food consumption. Economic Development and Cultural Change 53, Inter-American Development Bank, Evaluación cualitativa del uso de tarjetas electrónicas y cuentas de ahorro en la entrega de apoyos del programa oportunidades en zonas urbanas. Report. Karlan, D., Mobius, M., Rosenblat, T., Szeidl, A., Trust and social collateral. Quarterly Journal of Economics 124, Karlan, D., Zinman, J., Price and control elasticities of demand for savings. Working Paper. URL: Kast, F., Pomeranz, D., Saving more to borrow less: Experimental evidence from access to formal savings accounts in chile. Harvard Business School Working Paper URL: a8d cd5ae9702f03.pdf. La Porta, R., de Silanes, F.L., Zamarripa, G., Related lending. Quarterly Journal of Economics 118, Levy, S., Schady, N., Latin America's social policy challenge: Education, social insurance, redistribution. Journal of Economic Perspectives 27, Meyer, B., Mok, W.K.C., Sullivan, J.X., Household surveys in crisis. Journal of Economic Perspectives 29, Ozer, E.J., Fernald, L.C.H., Weber, A., Flynn, E.P., VanderWeele, T.J., Does alleviating poverty aect mothers' depressive symptoms? A quasi-experimental investigation of Mexico's Oportunidades programme. International Journal of Epidemiology 40, Parker, S.W., Teruel, G.M., Randomization and social program evaluation: The case of progresa. Annals of the American Acadamy of Political and Social Science 599, Prina, S., Banking the poor via savings accounts: Evidence from a eld experiment. Journal of Development Economics 115, Schaner, S., 2015a. The cost of convenience? transaction costs, bargaining, and savings account use in kenya. Working Paper. URL: ATM.pdf. Schaner, S., 2015b. Do opposites detract? Intrahousehold preference heterogeneity and inecient strategic savings. American Economic Journal: Applied Economics7,

47 Schultz, T.P., School subsidies for the poor: Evaluating the Mexican Progresa poverty program. Journal of Development Economics 74, Tobin, J., The interest-elasticity of transactions demand for cash. Review of Economics and Statistics 38, Wright, R., Tekin, E., Topalli, V., McClellan, C., Dickinson, T., Rosenfeld, R., Less cash, less crime: Evidence from the Electronic Benet Transfer program. NBER Working Paper URL: 46

48 Appendix Figure A1: Dierence between Treatment and Control in Marginal Propensity to Save Out of Transfer, Separated by Time with Account 0.6 (a) Wave 1 vs. Control (Older Accounts) 0.6 (b) Wave 2 vs. Control (Older Accounts) May 09 Sep 09 Jan 10 May 10 Sep 10 Jan 11 May 11 Sep 11 May 09 Sep 09 Jan 10 May 10 Sep 10 Jan 11 May 11 Sep (c) Wave 1 vs. Control (Younger Accounts) 0.6 (d) Wave 2 vs. Control (Younger Accounts) May 09 Sep 09 Jan 10 May 10 Sep 10 Jan 11 May 11 Sep 11 May 09 Sep 09 Jan 10 May 10 Sep 10 Jan 11 May 11 Sep 11 Sources: Administrative data from Banse on average account balances by bimester, transfer payments, and timing of card receipt. Notes: (a) N = 743,776 from 99,362 accounts; (b) N = 905,335 from 118,228 accounts; (c) N = 455,172 from 79,511 accounts; (d) N = 1,088,677 from 157,717 accounts. See the notes to Figure 7 for the specication. Accounts are split into older accounts and younger accounts based on the median account opening date, which is October 18,